quqxui / mpikgc

[LREC-COLING 2024] Multi-perspective Improvement of Knowledge Graph Completion with Large Language Models

Multi-perspective Improvement of Knowledge Graph Completion with Large Language Models

This is the code and data of paper Multi-perspective Improvement of Knowledge Graph Completion with Large Language Models, LREC-COLING 2024.

The entire project consists of three steps: (1) Generating data, (2) Processing data and (3) Runing KGC models.

Generating data

You can directly download the data we generated from Goolge Drive. Put the data in current directory, e.g., MPIKGC/LP_fb_wn_chatglm2/.

If you want to generate data with LLMs by yourself, you can run the following command to query LLMs for data:

First, cd generating_MPIKG/, then

For FB15k237:

python querying_llm_fb15k237.py --max_length 256 --temperature 0.2 --cuda 0 --batchsize 1 --LLMfold './../LP_fb_wn_llama2/' --LLMname ChatGLM2

For WN18RR:

python querying_llm_wn18rr.py --max_length 256 --temperature 0.2 --cuda 0 --batchsize 1 --LLMfold './../LP_fb_wn_llama2/' --LLMname ChatGLM2

For FB13 and WN11:

python querying_llm_TC_FB13_WN11.py --max_length 256 --temperature 0.2 --cuda 0 --batchsize 1 --LLMfold './../TC_fb_wn_llama2/' --LLMname ChatGLM2 --fb_or_wn FB13

After getting extracted keywords, you can genearate the top K matching entities for Structure Extraction, and change the default parameters in codes to obtain the desired data:

python textmatch4structure.py 

Finally, generating data in corresponding path, e.g., LP_fb_wn_chatglm2/FB15k237/:

cotdes.txt for MPIKGC-E

rel2des.txt for MPIKGC-R Global

rel2sentence.txt for MPIKGC-R Local

rel2reverse.txt for MPIKGC-R Reverse

struc/* for MPIKGC-S

Processing data

Merge the enhanced data with original KG to generate new KG with regular form:

cd processing_MPIKG
python merge.py 

Download KGC models from their repository: CSprom-KG, LMKE, SimKGC, KG-BERT. Put these project in current directory, e.g., MPIKGC/CSprom-KG/.

For diffrent KGC models, you need to slightly modify the form of data by runing:

python data4kgc_models.py 

This code will transfer the form of data to adapt to different KGC model requirements, and copy the data to corresponding model project folder.

Note that KG-BERT need same form as LMKE, You can directly copy it over.

Runing KGC models

Configure according to the environment and process requirements of each model. Hyperparameters for KGC models can be found in Appendices of our paper.

mpikgc's Issues

from src.read import read_ent2keyword

from src.read import read_ent2keyword

where is src.read in this repository??

How to map Freebase id to Wikipedia name?

Hello author, may I ask how to map freebase id to wikipedia name? For example the data in the file "FB15k_mid2name.txt", Thanks.

missing 1 required positional argument: 'top_loop'

我想知道在运行merge.py文件时候报错TypeError: merge_srcdes_srckeywords_rel2sentence_struc_srckeywords_top3_selfloop() missing 1 required positional argument: 'top_loop'是什么情况

merge.py ： The function on the last line lacks a parameter

Traceback (most recent call last):
File "merge.py", line 228, in
merge_srcdes_srckeywords_rel2sentence_struc_srckeywords_top3_selfloop('struc_srckeywords_top3_selfloop')
TypeError: merge_srcdes_srckeywords_rel2sentence_struc_srckeywords_top3_selfloop() missing 1 required positional argument: 'top_loop'